JPH0582956B2 - - Google Patents

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to improvements in speaker verification systems.

(Prior art) A speaker verification system verifies that the person speaking is the registered person by comparing the password input as voice with the voice pattern of the previously registered password and testing the degree of agreement. This is a system that determines whether Such a system is useful as a replacement for a seal in banking or as a replacement for a lock for entry security, and its realization is strongly desired.

Speaker confirmation systems that have been attempted in the past include, for example, 63 in Journal of the Acoustical Society of Japan, Vol. 35, No. 2 (February 1971).
"Speaker verification for telephone voice" from page to page 69
As described in a paper published under the title, a standard pattern created by each person's password is memorized and used. In this system, during the verification operation, a code (hereinafter simply referred to as a registration number) corresponding to the name claimed by the person to be verified and a password are uttered. Next, the standard pattern corresponding to the input registration number and the input password voice (hereinafter referred to as input pattern) are compared and the distance (evaluation value of the degree of difference between the patterns) is calculated. When this distance is smaller than a predetermined threshold, it is determined that the person to be verified is the real person, and conversely, when it is larger than the threshold, it is determined that this person is an impostor.

The problem with this type of speaker verification system is that
The problem is that it is difficult to set an appropriate threshold value.
In other words, even when the same person utters a fixed password, the voice pattern differs from time to time, and the degree of variation differs from person to person. For this reason, if the threshold is set too strictly (that is, to a small value), the person in question will often be judged to be the impostor, whereas if the threshold is set too loosely, the person in charge of the imposter will be incorrectly judged to be the person in question. .

A patent application has been made for a method of speaker verification with few such drawbacks.
It is described in specification No. 58-054132. The main points of the method are as follows. That is, in addition to the standard password pattern (hereinafter referred to as the "confirmation standard pattern"), another person's voice pattern similar to the confirmation standard pattern is registered as a "suppression standard pattern", and the result of matching with the input pattern is obtained. Only when the distance covered satisfies the following conditions (a) and (b),
It determines that the person is the person and issues a matching signal.

(B) Distance from the confirmation standard pattern < Distance from the suppression standard pattern (B) Distance from the confirmation standard pattern < Predetermined threshold According to this method, because of the judgment criterion (B),
The threshold value used in (b) can be set loosely. This reduces the number of erroneous determinations in which the person in question is the impostor, and also has the effect of preventing erroneous determinations in which another person is identified as the person in question (a).

(Problems to be Solved by the Invention) In the conventional method, a pattern of another person similar to a password confirmation standard pattern is selected as a suppression standard pattern. In this case, it is necessary to prepare patterns for many speakers as patterns for other people.
When targeting the general public, it has been difficult to prepare speech patterns for creating suppression standard patterns.

It is an object of the present invention to provide a suppression standard pattern creation method that eliminates the above-mentioned drawbacks, and to realize a speaker verification system that is free from operational problems.

(Means for Solving the Problem) The speaker verification system according to the present invention includes a verification standard pattern storage unit that stores verification standard patterns to be registered, and a vowel pattern storage unit that stores vowel patterns of multiple speakers prepared in advance. a suppression standard pattern synthesis section that connects the consonant part of the confirmation standard pattern and the vowel pattern to synthesize a suppression standard pattern; and a suppression that stores information for synthesizing a suppression standard pattern similar to the confirmation standard pattern. A designation section, an input pattern buffer that holds an input pattern uttered by the person to be verified, and a method based on the distance between this input pattern and the confirmation standard pattern, this input pattern, and the registered information input corresponding to this input pattern. A pattern matching unit that calculates the distance to the suppression standard pattern specified by the suppression designation unit, and whether or not the person to be matched is the person himself/herself using the distance to the confirmation standard pattern and the distance to the suppression standard pattern. and a determination unit that determines the.

(Operation) In the present invention, candidates for suppression standard patterns are synthesized in response to different passwords registered for each individual, and from among the candidates, a pattern that is similar to a password confirmation standard pattern is selected as a suppression standard pattern. . That is, as shown in Figure 2, the sequence of vowels that appeared in the confirmation standard pattern V _{i (1)} , V _{i (2)} , ..., V _{i (T)} ... (1) where T is the number of vowels. Based on this, the vowel part of the confirmation standard pattern is replaced with the vowel pattern of the majority of speakers. However, the speaker of the replaced vowel pattern is different from the speaker of the confirmation standard pattern. The vowel pattern sequence used for this synthesis is V ^m _i(1) , V ^m _i(2) , ..., V ^m _i(T) ; m≠n ...(2) (where n is the confirmation standard pattern utterance person). A plurality of suppression standard pattern candidates are generated. Pattern matching is performed between these suppression standard pattern candidates and the confirmation standard pattern to determine the distance. k selections are made from among the smaller distances, and the suppression standard pattern candidates to which the smaller distances are given are determined as suppression standard patterns. When checking,
The distance between the input pattern of the person to be matched and the confirmation standard pattern and the distance between the input pattern and the suppression standard pattern are determined, and based on the two distances, it is determined whether or not the person to be matched is the real person.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of the present invention. The vowel pattern storage unit 130 stores vowel patterns for a large number of people in advance. For example, vowel patterns such as /a/, /i/, /u/, /e/, /o/ are prepared for 1000 people. Let the set of these vowel patterns be {V ^m _i ; i=1, 2,..., 5: m=1, 2,...,
1000} ...(3).

First, the operation related to registration of the confirmation standard pattern will be explained. The registration information designation section 20 consists of a keyboard, which is used to input the information of the person to be registered, that is, the registration number n, and then press /
Enter the password fujisan/. Corresponding audio is also input from the microphone 10.
After this voice is analyzed by the analysis section 30, it is input to the input pattern buffer 40 (hereinafter referred to as input pattern
). This input pattern X is a confirmation standard pattern A ⁿ
It is sent to the confirmation standard pattern storage section 50 and stored in association with the registration number n.

Next, the synthetic information generation unit 100 receives the password vowel series V _{i (t)} (/
In the case of fujisan/, u, i, a) are received, and the vowel sequence V ^m _{i (t)} is sent to the suppression standard pattern synthesis unit 120 as synthesis information.

The suppression standard pattern synthesis unit 120 generates a vowel sequence V ^m _i(t)
Based on this, the vowel patterns V of speaker m are sequentially read out from the vowel pattern storage unit 130, and the vowel parts of the confirmation standard pattern are sequentially replaced. The pattern synthesized in this way is a suppression standard pattern candidate.
It is stored in the suppression standard pattern storage unit 90 as B ^m _o .

Thereafter, the composite information generation unit 100 sequentially changes the speaker number m from 1 to 1000. However, m≠n
Make it so that In response, the suppression standard pattern synthesis section 120 sequentially synthesizes the suppression standard pattern candidates B ^m _o , temporarily stored in the suppression standard pattern storage section 90, and sent to the pattern matching section 60 via the signal line b.

The pattern matching unit 60 performs a comparison process between the input pattern X and the suppression standard pattern B ^m _o sent from the suppression standard pattern storage unit 90 and sequentially calculates the distance D (X, B ^m _o ). calculate. In the suppression designation determination unit 70, the sequentially input distances D (X, B ^m _o ) are compared with each other, and the distance from the smallest one to, for example, the fifth smallest one is determined, and the corresponding number m is the suppression designation. The information is sent to section 80. Suppression specification section 8
0, these numbers m are associated with the registration number n and stored in the following format: m(n, 1), m(n, 2)...m(n, 5)...(4).

In this way, the confirmation standard pattern A ⁿ was obtained, and the preparation for suppression designation was completed as shown in equation (4). The above process is repeated while changing the registration number n every time a new person appears.

Next, the part related to the confirmation operation will be explained. The person to be verified operates the keys of the registration number designation section 20 to input the registration number n, which is registration information, and also inputs the password voice /fujisan/ from the microphone 10. The password voice is analyzed by the analysis section 30 in the same manner as at the time of registration, and is inputted as input pattern X to the input pattern buffer 40.

When the registration number n is input to the confirmation standard pattern storage section 50, a confirmation standard pattern A ⁿ is outputted accordingly. In response to this, the pattern matching section 60 first calculates the distance D(X, ^An ).

When the registration number n is given to the suppression designation unit 80 (4)
A suppression standard pattern designation signal m and registration number n of the expression are output to a signal line k. In response to this, the suppression standard pattern B ^m _o ;{m
=m(n, 1), m(n, 2), ..., m(n, 5)} are read out.

In the pattern matching section 60, the input pattern X is compared with these suppression standard patterns B ^m _o , distance D (X, B ^m _o ); m=m ₁ , m ₂ , . . .
m ₅ is calculated.

The determination unit 110 compares the distance D(X, A) with the distance group of equation (4), and finds that D(X, A) is smaller than any D(X, B ^m _o ), and D(X,
Only when A) is smaller than a predetermined threshold value, it is determined that the person to be verified is the real person.

Although the principle of the present invention has been explained above based on examples, these descriptions do not limit the scope of the present invention. In particular, in this specification, distance is used to compare voice patterns, but quantities with opposite magnitude relationships may also be used, such as correlation. In this case, the criterion (a)
It is self-evident that the inequality signs in and (b) are reversed. Further, instead of (a) as the determination criterion, distance to the suppression standard pattern>second threshold value can also be used. Furthermore, in the example, the vowel sequence of the confirmation standard pattern is inputted from the outside by the keyboard by the registered speaker in order to generate synthesis information for the suppression standard pattern candidate, but all combinations are synthesized assuming that the vowel sequence is unknown. A suppression standard pattern can also be obtained by selecting a pattern with a small distance by pattern matching.

Furthermore, in the suppression standard pattern synthesis unit 120,
The suppression standard pattern is created by directly connecting the vowel patterns, but in order to create a smooth connection near the connection point, various speech synthesis techniques are used, such as extrapolating the preceding and succeeding information and adding them with weights. You can also do it.

(Effects of the Invention) In the present invention, since the suppression standard pattern is synthesized from the vowel patterns of multiple speakers, it is possible to obtain the suppression standard pattern for any password. As a result, it is possible to realize a speaker verification system that has a higher verification rate than before and can accept any password.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is a diagram illustrating a method of creating a suppression standard pattern. In the figure, 10...microphone, 20...
Registration information designation section, 30... Analysis section, 40... Input pattern buffer, 50... Confirmation standard pattern storage section, 60... Pattern matching section, 70... Suppression designation determination section, 80... Suppression designation section, 90 ...Suppression standard pattern storage unit, 100...Synthesis information generation unit,
110... Determination unit, 120... Suppression standard pattern synthesis unit, 130... Vowel pattern storage unit, respectively.

Claims (1)

[Claims] 1. A confirmation standard pattern storage unit that stores confirmation standard patterns to be registered, a vowel pattern storage unit that stores vowel patterns of multiple speakers prepared in advance, and a consonant part of the confirmation standard pattern and the vowel pattern are connected. a suppression standard pattern synthesis unit that synthesizes a suppression standard pattern based on the confirmation standard pattern; a suppression specification unit that stores information for synthesizing a suppression standard pattern similar to the confirmation standard pattern; and an input pattern that holds an input pattern uttered by a person to be matched. Calculate the distance between the buffer, this input pattern, and the confirmation standard pattern, and the distance between this input pattern and the suppression standard pattern specified by the suppression specification section based on the registered information input corresponding to this input pattern. The method is characterized in that it includes a pattern matching section and a determination section that determines whether or not the person to be matched is the person himself/herself by using the distance to the confirmation standard pattern and the distance to the suppression standard pattern. Speaker verification system.